AI Signal 142
Geiger inventories every AI agent on a machine and reports what each can execute, read, or hold
Geiger is a read-only CLI tool that scans a machine for AI agents, MCP servers, plugins, and browser/editor extensions, then reports in plain language what each one can touch, code execution, filesystem access, network, or credentials.
As AI agent ecosystems proliferate across desktops and editors, the surface area of programs that can execute commands and hold secrets grows invisibly in dotfiles and config directories most people never inspect. Geiger gives engineers a single command to audit that surface, baseline it, and alarm on drift in CI or cron. The tool is read-only, telemetry-free, and reports secrets by shape only, making it safe to run on developer machines without exfiltrating anything.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Geiger detects agents across Claude Code, MCP hosts, CLI agents, editor extensions, and browser extensions, reporting capabilities like EXECUTES, HOLDS-SECRETS, BROAD-FILESYSTEM, and NETWORK for each finding.
A --strict flag exits non-zero if anything can execute code or hold secrets, and a --diff mode compares against a baseline JSON snapshot to alarm only on new or escalated findings.
The tool writes nothing except an explicitly named --json or --html output file, sends no telemetry, and redacts all credential values, reporting only key names and file paths.
THE READ
What the cluster adds up to.
Geiger addresses a visibility gap that has grown with the rapid spread of AI agent harnesses, MCP servers, and editor extensions. The tool runs as a single read-only command, `npx geiger-scan`, that reads configs and directories across multiple ecosystems without executing npm or any agent runtime. It inventories Claude Code agents and hooks, MCP servers in hosts like Claude Desktop, Cursor, Windsurf, VS Code, Cline, and Zed, CLI agents like Codex CLI, Gemini CLI, Aider, and others, plus AI extensions in browsers and editors. Each finding reports origin (registry, store, git, local script, remote server, or UNKNOWN-ORIGIN), capabilities, and an evidence path for manual verification.
The capability model is coarse but actionable: every finding is tagged with EXECUTES, HOLDS-SECRETS, BROAD-FILESYSTEM, BROAD-WEB, or NETWORK. This lets an engineer quickly see which agents on their machine can run code, which hold credentials, and which have broad filesystem or network reach. Geiger also recognizes policy wrappers, enforcement layers placed in front of MCP servers, and reports both the wrapper and the underlying server rather than hiding one behind the other. Hooks that fire without a prompt on events like UserPromptSubmit or PreToolUse are flagged separately, since they execute automatically.
For ongoing monitoring, Geiger supports a baseline-and-diff workflow. An engineer saves a reviewed inventory as JSON, then runs `geiger-scan --strict --diff baseline.json` in cron or CI. The strict exit code fires only when something new can execute code or hold secrets; the previously reviewed inventory stays quiet. This mirrors a lockfile mental model, accept what exists, alarm on change, and makes the tool practical for fleet-level drift detection rather than one-off auditing.
Adoption cost is minimal: no install, no account, no telemetry, and the tool never executes npm or any agent it discovers. The only write operation is an explicitly named --json or --html output file. Secrets are reported by shape only, key name, file, and secret type, with a redaction pass enforced by the test suite. The tool can also scan alternate home directories and project paths, which matters for incident response on mounted images or multi-user systems. Where it stops working is implicit: it reads configs and manifests, so agents that operate purely in memory or configure themselves at runtime outside of known config locations would not appear in a scan.
Written by elseif from the cluster below · checked for specifics the sources never containedTHE CLUSTER